Synchronize BF16 weight loads

[devshahofficial]

Summary

• Check CUDA errors while loading primary weights.
• Synchronize the BF16 cast on pufferl.default_stream before load_weights returns.
• Add a regression that guards cast launch checking and stream synchronization ordering.

Root cause

BF16 load_weights copied fp32 master weights, launched the fp32-to-BF16 cast on default_stream, and then returned without ordering that stream against subsequent inference streams. A first rollout could consume stale param_puf data.

Validation

• python -m pytest tests/test_cuda_load_weights.py tests/test_import_performance.py -q

GPU runtime validation still needs CUDA hardware; the local environment used for this patch does not include nvcc.

cc @Infatoshi for the original BF16 checkpoint repro.

Fixes #534